A seven-day gap between model releases is easy to dismiss as another version number. It is harder to dismiss when your evaluation process takes six weeks.

On 22 September 2026, OpenAI released GPT-6 Sol. On 29 September, GPT-6.1 Sol followed. One week. Google’s Gemini Flash family moved from 3.6 to 3.7 in 23 days, then to 3.8 in another 20. These dates do not tell us how long either model took to train. They tell us something more immediately useful: the frontier available to a team building a product can shift before that team finishes testing it.

I keep coming back to that mismatch. We plan around quarters. The model landscape increasingly moves within weeks. A sensible decision on Monday can need revisiting before the next steering committee.

The interesting question is no longer just “Which model is best?” It is “How do we build when the answer keeps changing?”

The pace, in five intervals

Here is a small, directly dated sample from two providers. It is deliberately narrow: recent announcements, with the days between them shown explicitly. A shorter release interval is evidence of a faster product cadence. It is not, by itself, evidence of a proportional jump in intelligence.

Weeks, then days

Calendar days between selected 2026 releases. Families and release types differ; these are not measurements of training duration. Sources: OpenAI changelog; Google’s 3.6, 3.7 and 3.8 announcements.

Zoom out: one frontier becomes many

Looking at one company understates the challenge. OpenAI, Anthropic, Google DeepMind, Meta, DeepSeek and xAI all contribute to the moving landscape. Releases overlap. Open-weight models open another route to deployment. A model can lose its advantage because a competitor improves, even if its own provider changes nothing.

The explorer below brings together the latest charts from my model timeline: release dates, same-provider intervals and eight benchmark views. Choose a company or a time window, then select a model to inspect the source and evaluation notes.

Explore the model timeline and benchmarks

69 selected events and 140 reported scores across six providers, as of 2 October 2026. Open the full interactive explorer ↗. This is a curated collection, not a complete release registry.

The selected timeline gets busier

Counts from the curated timeline, excluding one evaluation-only snapshot. 2026 is year-to-date through 2 October. Selection coverage changes over time: this chart describes this collection and cannot establish an unbiased industry-wide growth rate.

That last qualification matters. A crowded timeline is a useful picture of the environment we navigate. It is a poor substitute for a controlled measurement of progress. Counting a preview, a minor version and a new model family as one event each does not make them equivalent.

A higher score is only the beginning

Benchmark charts help, provided we resist the temptation to join every dot into a single story of intelligence. MMLU, HumanEval, GPQA Diamond, SWE-bench Verified and the AIME tests ask different questions. A score also depends on the prompt, reasoning budget, tools and evaluation harness.

The explorer keeps those tests separate. AIME 2024 and AIME 2025 have different panels. HLE results are not merged with HLE-Verified. Reported observations remain observations; a line between unlike evaluation setups would imply more comparability than the evidence supports.

As a test approaches its ceiling, small score changes tell us less. The practical question becomes whether a model can finish a harder task reliably, with an acceptable cost and enough traceability to check the result.

From answering questions to sustaining work

METR’s task-horizon research offers a different lens. Its 50% horizon estimates the human task duration at which an AI system succeeds half the time on the evaluated task distribution. In its updated TH1.1 dataset, the estimate rises from 3.5 minutes for GPT-4 to 320 minutes for Claude Opus 4.5. The uncertainty grows too.

Longer tasks within reach

Selected METR TH1.1 estimates; minutes on a logarithmic scale. Whiskers show reported 95% confidence intervals. Human task duration is not AI running time, and 50% success is not dependable unattended operation. Dataset and methodology.

METR estimates a roughly 131-day doubling time for its post-2023 TH1.1 trend, with a 107–161-day confidence interval. This is a fitted trend on a particular software-task suite, sensitive to task selection and modelling choices. It does not mean every kind of work doubles on that schedule.

Still, the direction is consequential. Better single answers are useful. Longer sequences of successful work change what can be delegated. The distance between those two is where much of the next product competition will happen.

Build for the replacement

For a financial data product, a digital-asset workflow or an internal research tool, I would treat the model as a component that will change. Keep a representative set of real tasks. Record the model version, instructions, tool access, cost and failure modes. Re-run that set when a credible replacement arrives.

A benchmark leaderboard cannot tell you whether a model silently mishandles a corporate action, confuses two token contracts or cites the wrong filing. Your own evaluation can. The metric worth defending is the cost of a correctly completed task, including human review and recovery when something goes wrong.

There is a useful discipline here: make upgrades cheap to test and easy to reverse. Otherwise, a faster frontier simply leaves you with more migration work.

When the research loop starts improving itself

There is another possible change in pace. Today, improved models help people build the next system. Recursive self-improvement, or RSI, would make the feedback loop more direct: a system improves the process that produces its own future improvements.

There are already bounded demonstrations. Sakana AI’s Darwin Gödel Machine modifies its agent codebase and tests the changes on coding benchmarks. That is improvement of the agent around a foundation model, not autonomous retraining of the foundation model itself. A September 2026 AIDE2 research paper reports seven successive improvements during an eight-day autonomous run, evaluated on held-out tasks. Neither result establishes an indefinitely accelerating research loop.

The distinction is central. A system improving a scaffold, a prompt or a search strategy is meaningful. Sustained improvement across research, experiments and the next model generation is a much stronger claim. A recent survey of RSI research makes this separation between bounded refinement and autonomous research loops explicit.

But suppose a future system reaches that stronger form of RSI, and each successful cycle increases its effective research productivity by a fixed proportion. The next cycle would start from a better base. Gains would compound.

If improvements compound

Hypothetical illustration, not observed capability or a forecast. All curves begin at 1× research productivity. The compounding curve uses P(n) = (1 + g)n; the linear comparison uses 1 + gn. The constrained case caps the same compounding path at an arbitrary 4×. Cycles have no assumed calendar duration. Neither the gain nor the cap is estimated from current systems.

Move the assumed gain from 10% to 30%. The first few steps look manageable. After repeated cycles, the difference is substantial. That is the mechanism behind an exponential scenario: not a giant leap on day one, but improvements that make subsequent improvements easier.

The smooth line hides the hardest part. An experiment can fail. An evaluator can reward the wrong thing. More capable research software still needs compute, data, energy and experiments that take physical time. Research on compute bottlenecks shows how strongly an intelligence-explosion scenario depends on assumptions about which inputs can substitute for one another.

Reaching RSI would therefore not guarantee exponential growth indefinitely. The important question would be whether successful cycles remain both productive and fast as the system becomes more capable. A graph of improvement cycles cannot answer how long those cycles take.

The model you evaluate today will not be the last one you use. If the research loop itself begins to compound, even our idea of a normal upgrade cycle may need replacing.